theislab/scarches

scPoli Model for Unsupervised Use

Opened this issue · 1 comments

Version history

scarches: 0.5.10

Issue

When attempting to initialize the scPoli model for unsupervised learning, the constructor seems to require a label_indices=[] argument to be passed as an empty list, which should be optional in unsupervised scenarios.

Hi,
To use scPoli in unsupervised settings, please note that you should:

  1. Set all values in the passed cell_type_keys to a dummy value (e.g. 'Unknown') and pass the unknown_ct_names=['Unknown'] parameter to scPoli.
  2. Additionally, Set labeled_indices=[]

The second requirement may be relaxed in an upcoming merge request.
Please note that scPoli without cell-type information is equivalent to trVAE with batches handled using embedding vectors rather than one-hot encodings.